From: Matthew Dillon <dillon@apollo.backplane.com>
Date: Mon, 12 Jan 2009 21:46:20 +0000 (-0800)
Subject: Set the auto-reflush flag if the flusher is unable to completely flush
X-Git-Tag: v2.3.0~43
X-Git-Url: https://gitweb.dragonflybsd.org/dragonfly.git/commitdiff_plain/5c667a24d6f71ecb8735c064088ba0a342a05cc8

Set the auto-reflush flag if the flusher is unable to completely flush
an inode due to dependencies or memory considerations.

This should hopefully solve reported 'vinvalbuf: dirty bufs' panics,
which occur when VOP_FSYNC() returns before the inode is completely
flushed.

Reported-by: "Justin C. Sherrill" <justin@shiningsilence.com>
---

diff --git a/sys/vfs/hammer/hammer_inode.c b/sys/vfs/hammer/hammer_inode.c
index e8df385ad4..537bd19487 100644
--- a/sys/vfs/hammer/hammer_inode.c
+++ b/sys/vfs/hammer/hammer_inode.c
@@ -2048,6 +2048,14 @@ hammer_flush_inode_done(hammer_inode_t ip, int error)
 
 	hmp = ip->hmp;
 
+	/*
+	 * Auto-reflush if the backend could not completely flush
+	 * the inode.  This fixes a case where a deferred buffer flush
+	 * could cause fsync to return early.
+	 */
+	if (ip->sync_flags & HAMMER_INODE_MODMASK)
+		ip->flags |= HAMMER_INODE_REFLUSH;
+
 	/*
 	 * Merge left-over flags back into the frontend and fix the state.
 	 * Incomplete truncations are retained by the backend.